chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[Test] Refactor KL divergence and prefix cache branching to kits (#19715)

Browse Source
This commit is contained in:
roikoren755
2026-03-12 16:11:59 +08:00
committed by GitHub
parent 46b558445d
commit 067353f67b
8 changed files with 244 additions and 552 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/4-gpu-models/test_qwen3_next_models.py
+21 -117
View File
@@ -1,129 +1,33 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=350, suite="stage-c-test-4-gpu-h100")
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3Next(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
class TestQwen3Next(
GSM8KMixin, KLDivergenceMixin, PrefixCacheBranchingMixin, DefaultServerBase
):
model = QWEN3_NEXT_MODEL
cache_chunk_size = 64
gsm8k_accuracy_thres = 0.93
kl_div_thres = 0.0025
other_args = [
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
]
if __name__ == "__main__":
test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
+85 -275
View File
@@ -1,299 +1,109 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.kits.prefix_cache_branching_kit import PrefixCacheBranchingMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=500, suite="stage-c-test-4-gpu-h100")
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
# MTP has higher KL divergence threshold
ACC_THRESHOLDS_MTP = {
QWEN3_NEXT_MODEL: {"kl_div": 0.008, "gsm8k": 0.93},
}
class TestQwen3NextMTP(GSM8KMixin, KLDivergenceMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL
gsm8k_accuracy_thres = 0.93
kl_div_thres = 0.0025
other_args = [
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"no_buffer",
"--disable-radix-cache",
]
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3NextMTPTopk(
GSM8KMixin, KLDivergenceMixin, PrefixCacheBranchingMixin, DefaultServerBase
):
model = QWEN3_NEXT_MODEL
cache_chunk_size = 64
gsm8k_accuracy_thres = 0.93
kl_div_thres = 0.008
other_args = [
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"8",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
]
class TestQwen3NextMTP(CustomTestCase):
# TODO(hzh): After merging the PR that fixes specv2 to correctly return log probs,
# add KLDivergenceMixin back. https://github.com/sgl-project/sglang/pull/18645
class TestQwen3NextMTPV2(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL
gsm8k_accuracy_thres = 0.93
other_args = [
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
]
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"no_buffer",
"--disable-radix-cache",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
class TestQwen3NextMTPTopk(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"8",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS_MTP[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
class TestQwen3NextMTPV2(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
envs.SGLANG_ENABLE_SPEC_V2.set(True)
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
super().setUpClass()
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
# TODO(hzh): After merging the PR that fixes specv2 to correctly return log probs, re-open the tests below. https://github.com/sgl-project/sglang/pull/18645
# def test_input_output_logprobs_match(self):
# test_input_output_logprobs_match_helper(
# self.base_url,
# ACC_THRESHOLDS,
# self.model,
# max_samples=32,
# max_new_tokens=512,
# )
# def test_input_output_logprobs_match_prefill_cache_hit(self):
# test_input_output_logprobs_match_prefill_cache_hit_helper(
# self.base_url,
# ACC_THRESHOLDS,
# self.model,
# max_samples=32,
# max_new_tokens=512,
# )
# def test_input_output_logprobs_match_decode_cache_hit(self):
# test_input_output_logprobs_match_decode_cache_hit_helper(
# self.base_url,
# ACC_THRESHOLDS,
# self.model,
# max_samples=32,
# max_new_tokens=512,
# )
super().tearDownClass()
if __name__ == "__main__":
test/registered/models/test_qwen3_next_models_fp4.py
+18 -55
View File
@@ -1,70 +1,33 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.srt.utils import get_device_sm
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=500, suite="nightly-4-gpu-b200", nightly=True)
QWEN3_NEXT_MODEL_FP4 = "nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL_FP4: {"kl_div": 0.0025, "gsm8k": 0.93},
}
@unittest.skipIf(
get_device_sm() < 100, "Test requires CUDA SM 100 or higher (Blackwell)"
)
class TestQwen3NextFp4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL_FP4
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--quantization",
"modelopt_fp4",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
class TestQwen3NextFp4(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL_FP4
gsm8k_accuracy_thres = 0.93
other_args = [
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--quantization",
"modelopt_fp4",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
]
if __name__ == "__main__":
test/registered/models/test_qwen3_next_models_pcg.py
+9 -48
View File
@@ -3,17 +3,10 @@ Qwen3 Next piecewise CUDA graph tests.
"""
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(
est_time=400,
@@ -22,46 +15,14 @@ register_cuda_ci(
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
class TestQwen3NextPiecewiseCudaGraph(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp",
"4",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
class TestQwen3NextPiecewiseCudaGraph(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL
gsm8k_accuracy_thres = 0.93
other_args = [
"--tp",
"4",
]
if __name__ == "__main__":
test/registered/radix_cache/test_swa_radix_cache_kl.py
+13 -57
View File
@@ -1,69 +1,25 @@
import unittest
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.kl_divergence_kit import KLDivergenceMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
MODEL = "openai/gpt-oss-20b"
ACC_THRESHOLDS = {
MODEL: {"kl_div": 0.002},
}
register_cuda_ci(est_time=100, suite="stage-b-test-large-1-gpu")
class TestSWARadixCacheKL(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
# Use a lower mem-fraction-static to avoid OOM during input logprobs
# gathering. With PCG enabled, more memory is reserved for CUDA graph
# captures, so the static fraction should be lower.
other_args=[
"--tp-size",
"1",
"--mem-fraction-static",
"0.70",
"--disable-piecewise-cuda-graph",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=2048,
)
class TestSWARadixCacheKL(KLDivergenceMixin, DefaultServerBase):
model = MODEL
kl_div_thres = 0.002
kl_div_decode_max_new_tokens = 2048
other_args = [
"--tp-size",
"1",
"--mem-fraction-static",
"0.70",
"--disable-piecewise-cuda-graph",
]
if __name__ == "__main__":